# Mechanism figure sources

These illustrations explain source-reported mechanisms. The diagrams, lettering and vector icons were newly drawn for the review. Source images in the TissueLab, MACRO and ReVision3D panels retain their original pixels and annotations.

## medexpmem
Two-phase experience construction in MedExpMem. Phase I diagnoses educational cases without memory access and extracts pairwise notes from errors using the reference diagnosis, expert findings and discussion. The central memory indexes notes by department, organ and differential pair. Phase II re-diagnoses the same cases with memory access and supplements or merges insights from errors. For new cases, anatomical scoping restricts pair-key embedding retrieval, case-aware filtering removes irrelevant notes, and retained confusions, discriminators and decision rules support reasoning. No model-parameter update or test-time memory-writing path is depicted. A vs. B, M vs. N and the illustrated hierarchy are schematic examples from the source, with no invented scores or patient results. Adapted from MedExpMem, Fig. 1, Table 1 and sections 2.2 to 2.4 (arXiv:2605.22872v1). All graphics and lettering are newly drawn vectors. 

## macro
Composite-tool discovery in MACRO. A verified training trajectory supplies tool calls for the memory and discovery mechanism. The illustrated glaucoma workflow resizes the image, converts it to grayscale, segments the optic disc and cup, and computes the cup-to-disc ratio (CDR). Matching highlights show an illustrative contiguous fragment recurring across successful trajectories. A candidate is registered when its frequency exceeds the threshold τ. The resulting composite becomes a callable action alongside the primitive tools. The sequence rows are schematic and do not report measured occurrence counts. Redrawn from MACRO, Fig. 2, §3.3 and the glaucoma example in Fig. 4 (arXiv:2603.05860v1), CC BY 4.0. The small fundus thumbnail is cropped from the source Fig. 4; all operator icons, lettering and paths are editable vectors.

## tissuelab
Expert-guided active learning in TissueLab. Predictions are reviewed, cell annotations are retained as supervision, a NuClass-based classifier is updated, and analysis is rerun for further review. The upper-left and lower-left crops are the source colon example at the initial iteration and after 30 minutes. The upper-right crop illustrates the actual cell-label interface from Supplementary Fig. 1c; it is a different field of view and is not a matched intermediate image of the left-hand example. The retained-label and classifier cards are schematic. The cycle expresses the reported interactive mechanism, not a new experiment. Adapted from A co-evolving agentic AI system for medical imaging analysis, §2.1, §2.5, the unnumbered illustration d–e, and Supplementary Fig. 1c (arXiv:2509.20279v1), CC BY 4.0. Source pixels and overlays are unchanged apart from cropping and proportional display.

## diagagent
Reward-driven policy evolution in DiagAgent. An initial inquiry seeds diagnostic rollouts containing examination requests, simulated results and a final diagnosis (Dx). DiagGym generates results conditioned on the case profile and examination history. Rollouts are evaluated using diagnosis match, examination-set F1 and termination within the 12-turn limit. The displayed weights are 1, 0.5 and 1 as reported in §5.2. GRPO updates the agent parameters through the return path. The three rows are schematic and do not represent new patient records, measured trajectories or the rollout count used in training. Adapted from Evolving Interactive Diagnostic Agents in a Virtual Clinical Environment, Fig. 7b and §5.2 (arXiv:2510.24654v2), CC BY 4.0. All icons, shapes, text and paths are newly drawn editable vectors. No source image pixels are included.

## ecgrsi
Arrow revision of the all-vector ECG-RSI data-factory draft. The three synthesis arrows have visible shafts and consistent heads. The rejection and retry arrows have matching clearance from the frames, and retry aligns with the first column. The cropped external arrow beside the planner is omitted because its neighbouring source panel is absent. All lettering, icons and frames retain their preceding positions. No bitmap images are embedded in the PDF or SVG. The source waveform is an illustration from the original framework figure, not a new patient record or result. Adapted from ECG-RSI, Fig. 1, arXiv:2610.17718v1. The source paper is deposited under arXiv non-exclusive distribution terms, rather than a CC licence. Source attribution is retained, and no additional licence is assigned to third-party material here. 

## revision3d
Reading-pipeline revision in ReVision3D. CT replay and a fixed mask-grounded verifier support attribution of missed lesions. A fixed LLM designer and an outcome playbook support proposals for acquisition, perception, training recipe or inference. Replay can screen views, rewards or saved outputs, while perception revisions require training. Candidates are evaluated under the fixed development objective J. Outcomes feed back to the designer. Retained revisions update the reading system. Solid blue arrows indicate the forward flow, and gray dashed arrows indicate feedback and retained updates. The routing rule was formulated retrospectively and did not schedule the reported experiments. The categories and decorative symbols have no quantitative scale. Adapted from ReVision3D, Fig. 1 and section 3 (arXiv:2609.32984v2), CC BY 4.0. CT thumbnails are unaltered rectangular crops of the original figure, including its annotations. All other icons, lettering, frames and connectors are newly drawn vectors. No new patient data or measured results are depicted.
